Nebius Buys a Startup That Kills 'Idle GPU Tax' — Then Raises Its Own GPU Prices the Same Day
Nebius Group (NASDAQ: NBIS) announced October 1 that it acquired Inferize, a 17-person Tel Aviv startup barely ten months old, to fold into Token Factory, its managed inference platform. Nebius did not disclose a price. Israeli outlets CTech and Globes independently pegged the deal at $100 million to $150 million for a company whose public footprint, as of the announcement, was still a waitlist landing page. Inferize's pitch is narrow and technical: a system that snapshots a warmed serving engine's full CPU and GPU state — compiled kernels, CUDA graphs included — and restores it in seconds rather than the roughly 23 minutes a cold GPU normally takes to come back online, a claim based on the company's own single-vendor benchmark against DeepSeek V4 Pro with no independent verification yet published.
The founders, Guy Bortnikov and Lior Gorbonos, previously built Granulate, a workload-optimization startup Intel acquired in 2022 — giving Nebius a repeat team rather than an unproven one. That pedigree matters because this is not Nebius's first purchase in this exact lane. Eigen AI, bought for a disclosed $643 million, and Clarifai, whose compute-orchestration team and technology Nebius absorbed for roughly $97.4 million, both landed inside Token Factory within the prior five months. Stacking the three deals puts Nebius's inference-efficiency shopping spree north of $840 million since Token Factory launched in November 2025 — a notably large bet on cold-start latency and GPU utilization as the next competitive front in AI cloud, ahead of CoreWeave, Crusoe and Together AI.
The same-day price hike
The complicating fact missing from Nebius's own announcement: its public rate card moved upward the same day. H200 instances went from $4.50 to $5.40 an hour, H100 from $3.85 to $4.50, and B200 from $7.15 to $8.50, according to a rate-card comparison published alongside the deal coverage. On an eight-GPU H200 minimum replica, that is roughly $31,500 a month versus about $26,300 before — a $5,200 monthly increase landing on customers' bills the same day Nebius pitched the acquisition as eliminating an 'idle GPU tax.' An efficiency acquisition that reprices the vendor's own costs rather than the customer's rate card is not quite the same thing as the savings story in the press release, and nothing in Nebius's announcement connects the two moves or explains the timing.
"Removing that cost is what we built Inferize to do, and Nebius is where it can go straight into the platform." — Guy Bortnikov, co-founder and CEO, Inferize
There's also a gap between what Inferize's technology actually fixes and what customers need fixed. Fast state-restore only solves half of the cold-start problem: Nebius's own billing policy still warns that once GPU capacity is released, 'it may not be immediately available again.' Minimum-replica customers reserve both a warm serving engine and a physical GPU slot; Inferize's snapshot-and-restore addresses the warm-engine half, not the guarantee that a GPU slot is still there to restore onto. Nebius has not published an integration timeline or a shipped feature for Token Factory customers, continuing a pattern from the Eigen AI and Clarifai deals, where neither acquisition changed customer-facing contract terms at the time of announcement.
Why the spree, and what it costs
The acquisitions make more sense against Nebius's Q2 2026 numbers, released in August: revenue of $582.3 million, up 454% year over year, with first-half revenue at $981.3 million, up 529%. But the growth is bought, not earned yet — Q2 net loss from continuing operations was $190.4 million (adjusted net loss narrowed to $33.2 million, down from $91.5 million a year earlier), while first-half capital expenditure hit $8.13 billion, up 671% year over year. Cash on hand more than doubled in six months, from $3.68 billion to $8.04 billion, giving Nebius the balance sheet to keep buying inference teams even as it posts nine-figure quarterly losses. Three acquisitions in five months, each folded quietly into Token Factory with no disclosed full terms and no customer-facing commitments, reads less like opportunistic deal-making and more like Nebius racing to own the inference-efficiency layer before a hyperscaler or a well-funded rival (xAI, CoreWeave, or OpenAI's own infrastructure buildout) does it first — funded, for now, by GPU renters rather than by the acquisitions paying for themselves.